#8157: Add golden function

tests/sweep_framework/sweeps/eltwise/unary/leaky_relu/leaky_relu.py

@@ -3,32 +3,32 @@
 # SPDX-License-Identifier: Apache-2.0
 
 from typing import Optional, Tuple
+from functools import partial
 
 import torch
 import random
 import ttnn
 
 from tests.ttnn.utils_for_testing import check_with_pcc, start_measuring_time, stop_measuring_time
+from tests.sweep_framework.utils import gen_shapes
+from tests.tt_eager.python_api_testing.sweep_tests.generation_funcs import gen_func_with_cast_tt
+from models.utility_functions import torch_random
 
 # Override the default timeout in seconds for hang detection.
 TIMEOUT = 30
 
-random.seed(0)
-
 
 # Parameters provided to the test vector generator are defined here.
 # They are defined as dict-type suites that contain the arguments to the run function as keys, and lists of possible inputs as values.
 # Each suite has a key name (in this case "suite_1") which will associate the test vectors to this specific suite of inputs.
 # Developers can create their own generator functions and pass them to the parameters as inputs.
 parameters = {
-    "nightly": {
-        "batch_sizes": [(1,)],
-        "height": [384, 1024],
-        "width": [1024, 4096],
-        "dtype": [ttnn.bfloat16, ttnn.bfloat8_b],
-        "input_memory_config": [ttnn.DRAM_MEMORY_CONFIG],
+    "leaky_relu11": {
+        "input_shape": gen_shapes([1, 1, 32, 32], [6, 12, 256, 256], [1, 1, 32, 32], 16),
+        "input_a_dtype": [ttnn.bfloat16, ttnn.bfloat8_b],
+        "input_a_layout": [ttnn.TILE_LAYOUT, ttnn.ROW_MAJOR_LAYOUT],
+        "input_a_memory_config": [ttnn.DRAM_MEMORY_CONFIG],
         "output_memory_config": [ttnn.DRAM_MEMORY_CONFIG],
-        "layout": [ttnn.TILE_LAYOUT, ttnn.ROW_MAJOR_LAYOUT],
         "negative_slope": [-0.5, 0, 0.01, 0.5],
     },
 }
@@ -38,7 +38,7 @@
 # If invalidated, the vector will still be stored but will be skipped.
 # Returns False, None if the vector is valid, and True, str with a reason for invalidation if it is invalid.
 def invalidate_vector(test_vector) -> Tuple[bool, Optional[str]]:
-    if test_vector["layout"] == ttnn.ROW_MAJOR_LAYOUT:
+    if test_vector["input_a_layout"] == ttnn.ROW_MAJOR_LAYOUT:
         return True, "Row Major layout is not supported"
     return False, None
 
@@ -48,28 +48,35 @@ def invalidate_vector(test_vector) -> Tuple[bool, Optional[str]]:
 # The runner will call this run function with each test vector, and the returned results from this function will be stored.
 # If you defined a device_mesh_fixture above, the object you yielded will be passed into this function as 'device'. Otherwise, it will be the default ttnn device opened by the infra.
 def run(
-    batch_sizes,
-    height,
-    width,
-    dtype,
-    input_memory_config,
+    input_shape,
+    input_a_dtype,
+    input_a_layout,
+    input_a_memory_config,
     output_memory_config,
-    layout,
     negative_slope,
     *,
     device,
 ) -> list:
-    input_shape = (*batch_sizes, height, width)
-
-    torch_input_tensor = torch.randn(input_shape, dtype=torch.float32)
-    torch_output_tensor = torch.nn.functional.leaky_relu(torch_input_tensor, negative_slope)
-
-    input_tensor = ttnn.from_torch(
-        torch_input_tensor, dtype=dtype, device=device, memory_config=input_memory_config, layout=layout
+    data_seed = random.randint(0, 20000000)
+    torch.manual_seed(data_seed)
+
+    torch_input_tensor_a = gen_func_with_cast_tt(
+        partial(torch_random, low=-100, high=100, dtype=torch.float32), input_a_dtype
+    )(input_shape)
+
+    golden_function = ttnn.get_golden_function(ttnn.leaky_relu)
+    torch_output_tensor = golden_function(torch_input_tensor_a, negative_slope=negative_slope)
+
+    input_tensor_a = ttnn.from_torch(
+        torch_input_tensor_a,
+        dtype=input_a_dtype,
+        layout=input_a_layout,
+        device=device,
+        memory_config=input_a_memory_config,
     )
 
     start_time = start_measuring_time()
-    result = ttnn.leaky_relu(input_tensor, negative_slope, memory_config=output_memory_config)
+    result = ttnn.leaky_relu(input_tensor_a, negative_slope, memory_config=output_memory_config)
     output_tensor = ttnn.to_torch(result)
     e2e_perf = stop_measuring_time(start_time)
 

tests/ttnn/sweep_tests/sweeps/sweeps/leaky_relu.py

@@ -0,0 +1,55 @@
+# SPDX-FileCopyrightText: © 2023 Tenstorrent Inc.
+
+# SPDX-License-Identifier: Apache-2.0
+
+from typing import Optional, Tuple
+
+import torch
+import torch.nn.functional as F
+
+import ttnn
+
+from tests.ttnn.utils_for_testing import check_with_pcc
+from models.utility_functions import torch_random
+
+
+parameters = {
+    "batch_sizes": [(1,)],
+    "height": [384, 1024],
+    "width": [1024, 4096],
+    "input_dtype": [ttnn.bfloat16],
+    "input_memory_config": [ttnn.DRAM_MEMORY_CONFIG],
+    "output_memory_config": [ttnn.DRAM_MEMORY_CONFIG],
+    "layout": [ttnn.TILE_LAYOUT],
+    "negative_slope": [-0.5, 0, 0.01, 0.5],
+}
+
+
+def run(
+    batch_sizes,
+    height,
+    width,
+    input_dtype,
+    input_memory_config,
+    output_memory_config,
+    layout,
+    negative_slope,
+    *,
+    device,
+) -> Tuple[bool, Optional[str]]:
+    input_shape = (*batch_sizes, height, width)
+
+    low = -100.0
+    high = 100.0
+
+    torch_input_tensor = torch_random(input_shape, low, high, dtype=torch.float32)
+    torch_output_tensor = F.leaky_relu(torch_input_tensor, negative_slope)
+
+    input_tensor = ttnn.from_torch(
+        torch_input_tensor, dtype=input_dtype, device=device, layout=layout, memory_config=input_memory_config
+    )
+
+    output_tensor = ttnn.leaky_relu(input_tensor, negative_slope, memory_config=output_memory_config)
+    output_tensor = ttnn.to_torch(output_tensor)
+
+    return check_with_pcc(torch_output_tensor, output_tensor, 0.999)

ttnn/cpp/ttnn/operations/eltwise/unary/unary_pybind.hpp

@@ -1576,7 +1576,6 @@ void py_module(py::module& module) {
 
         )doc");
 
-
     detail::bind_unary_operation_with_float_parameter(module, ttnn::leaky_relu, "slope", "The slope parameter for the Leaky ReLU function", "",
         R"doc(Supported dtypes, layouts, and ranks:
 

ttnn/ttnn/operations/unary.py

@@ -256,6 +256,15 @@ def _golden_function_relu_min(input_tensor_a, *args, lower_limit, **kwargs):
     return torch.max(input_tensor_a, torch.tensor(lower_limit))
 
 
+def _golden_function_leaky_relu(input_tensor_a, *args, negative_slope, **kwargs):
+    import torch
+
+    return torch.nn.functional.leaky_relu(input_tensor_a, negative_slope)
+
+
+ttnn.attach_golden_function(ttnn.leaky_relu, golden_function=_golden_function_leaky_relu)
+
+
 ttnn.attach_golden_function(ttnn.relu_min, golden_function=_golden_function_relu_min)